Automated confidence intervals that achieve nominal coverage without increasing length relative to rdrobust

Determine whether one can design automated end-to-end procedures for constructing confidence intervals in regression discontinuity designs that consistently attain nominal coverage across realistic empirical settings while producing interval lengths comparable to those from the Calonico–Cattaneo–Titiunik (2014) rdrobust method.

Background

The authors compare rdrobust (bias-corrected) and RDHonest (bias-aware) methods on simulations calibrated via WGANs to landmark RDD applications. They find that rdrobust often undercovers while RDHonest, though valid, yields substantially wider intervals.

These findings motivate a pragmatic question about whether an automated approach can reliably achieve nominal coverage without the cost in width observed for RDHonest. The authors present PLRD as providing an affirmative answer in their numerical studies, but the question is posed explicitly as open prior to introducing their method.